Anthropic's second-quarter federal lobbying disclosure names export controls and model distillation repeatedly and by statute. OpenAI's, filed for the same period, does not mention China, open weights, export controls, or distillation at all. The contrast lands one day after the New York Times reported, on anonymous sources, that both companies privately pressed Washington to restrict Chinese open-weight models - a claim neither filing confirms and neither company has denied on the record.
Key facts
- Anthropic reported $1,970,000 in in-house lobbying expense for the second quarter of 2026, filed July 20.
- Its disclosure lists "Matters related to Export Controls and Distillation" and separately "Matters related to Artificial Intelligence Distillation," plus the AI Diffusion Rule and the AI OVERWATCH Act.
- OpenAI reported $1,200,000 for the same quarter, with issue descriptions limited to artificial intelligence, cloud computing and infrastructure, cybersecurity, copyright and privacy.
- Primary sources: the Anthropic LD-2 filing and the OpenAI LD-2 filing, both public on the Senate's lobbying disclosure system.
Some background on what these documents are. Under the Lobbying Disclosure Act, any organisation lobbying the federal government files a quarterly report naming the general issue areas, the specific bills and rules it worked on, which chambers and agencies it contacted, and roughly what it spent. The forms do not record positions. A company that lobbied to loosen export controls and one that lobbied to tighten them file identical-looking lines. So a filing is evidence of subject matter and attention, never of intent.
With that caveat firmly in place, the subject matter is striking. Anthropic's report contacted the House, the Senate, the White House Office, the Commerce Department and the Treasury, and its issue list runs to seventeen distinct items, including BIS-2025-0001, the Artificial Intelligence Diffusion Rule and export controls; H.R.6875, the AI OVERWATCH Act, described as "matters related to export controls and distillation"; H.R.8170, the MATCH Act; and S.1705, the Chip Security Act. Distillation appears twice as its own standalone issue. OpenAI's, by contrast, is three short lines of generic categories across its issue pages.
Distillation is the technical idea sitting underneath all of this, and it is worth understanding. Distillation means training a smaller, cheaper model on a larger model's outputs, so the student learns to imitate the teacher without ever seeing the teacher's weights. It is a completely standard technique that labs routinely apply to their own models. The contested version is doing it to somebody else's model at industrial scale through the public API, which is what Anthropic alleges in its published account of distillation attacks, naming fake-account campaigns it attributes to Chinese labs. Anthropic's argument is that the copy inherits the capabilities without the safety training - and that open-sourcing the result multiplies the risk. That framing is a company assertion, not an adjudicated finding, and it sits alongside earlier reporting on alleged distillation of Anthropic's models and floated sanctions.
The New York Times story, "Silicon Valley Splits Over Closing the Borders to Chinese A.I." by Mike Isaac, Kate Conger, Ana Swanson and Meaghan Tobin, is narrower than the headlines it generated. It attributes the private-lobbying claim to five people close to the discussions, gives no bill, rule, meeting or threshold, and reports that officials seemed more likely to handle models individually as national-security cases than to impose a blanket ban. Axios has the strongest on-record counterweight, quoting an OpenAI spokesperson saying advances in Chinese open-weight models are "not an argument against openness." That is not a denial that OpenAI lobbied; it is a statement about openness in general.
OpenAI's public position is genuinely conditional rather than hostile. It signed the July 24 open-weights letter, which argues open weights should not be prohibited and that unlawful extraction should be handled through targeted legal mechanisms. It has also proposed mandatory pre-release evaluation by the US AI safety agency for the most capable frontier models - a real gate, but one its own blueprint says should recommend mitigations rather than approve or block releases, and which lets developers deploy if the agency misses its deadline. Those two positions are compatible. They are not the same as wanting Chinese models banned.
Community reaction has drifted toward "everyone versus Anthropic," which the record does not support: the allegation covers both labs, and it is Anthropic that has published its policy reasoning at length while OpenAI's own filings say almost nothing. The best available framing of the stakes comes from open-weights advocate Tobi Knaup, who argues that weights turn a model into a platform for quantizations, fine-tunes and serving stacks, so restricting them would cut American builders out of a compounding ecosystem rather than merely denying them a chatbot.
The honest caveat is the central one: filings show attention, not asks. No LD-2, meeting readout, or named participant substantiates the specific allegation that either company requested restrictions on Chinese open-weight models. What the filings do establish is that one of these two companies has been telling the federal government, in writing, that distillation and export controls are its business - and the other has not.
Originally published on Ground Truth, where every claim is checked against the primary source.
Top comments (0)